India bets on ‘sealed rooms’ to train AI on sensitive data without exposing it

New Delhi/Bengaluru: India on Friday showcased a system that allows companies and researchers to train artificial intelligence (AI) algorithms on sensitive data held by hospitals, banks and government agencies without the information leaving those institutions or being exposed to them.

The Data Empowerment and Protection Architecture (DEPA) for AI Training framework was launched by the ministry of electronics and information technology (MeitY) and developed primarily by iSPIRT, a volunteer-led technology think-tank that has contributed to the development of India’s Unified Payments Interface (UPI) and India Stack, a set of digital public infrastructure systems that enable services such as digital identity (Aadhaar), payments and the secure sharing of data (DigiLocker).

Depa’s central idea is to bring AI models to the data, rather than move the data to researchers, a senior government official said at the framework’s unveiling in Bengaluru on Friday.

To be sure, Depa itself is not new. In 2020, NITI Aayog, the government’s top policy think-tank, introduced it as a framework that allows citizens to securely share their data with their consent. The new framework, developed by iSPIRT, extends this approach to training AI models on sensitive data belonging to millions of people without requiring the data to leave the institutions that hold it.

Consultations on Depa for AI will begin in the coming weeks, before it is brought under a legal framework, a process which could take years, the stakeholders confirmed. Depa comes as India prepares to implement its first data privacy law, the Digital Personal Data Protection (DPDP) Act, 2023, in the coming months. Meanwhile, several lawsuits involving AI platforms and their use of proprietary data remain sub judice.

The official cited earlier called the approach “a democratic third way” between the US, where companies buy data in commercial deals—and China, where the state pools it.

Industry stakeholders called the framework potentially crucial for effective training of domain-specific AI models, applications and products.

“While the framework could be potentially vital, the real use cases could be for startups like Sarvam and others seeking to build domain-specific AI for healthcare, education, critical communications and others. But, the crucial aspect would be to build a trust framework before that, and also validating the quality of vast tranches of sensitive data—both of which are difficult to establish,” said Jibu Elias, India lead for responsible computing at global open-source internet body, Mozilla Foundation.

Under Depa, data remains with its owner, while a “sealed room”, which is a secure computing environment, runs only what an authorised digital contract permits. No one can access the data inside, not even the company operating the system, according to the launch presentation. A developer ssends in an AI model and receives only the trained model or, in cases such as a loan application, the decision itself.

On Friday, early pilots of the framework were demonstrated. In one such trial, a bank that has been testing the approach since March 2026 identified which customers of a fintech qualified for its credit cards before seeking anyone’s consent, and then requested permission only from those who had a real offer. The pilot handled more than 2 million requests in two weeks. In another pilot, hospitals and a health-tech startup are training a tumour-detection model on MRI scans the startup never sees.

Ajay K Sood, principal scientific adviser to the government, Meity secretary S. Krishnan, iSpirt co-founder Sharad Sharma, and representatives from HDFC Bank, Sarvam and Emergent AI, among others, attended the event.

The framework, the stakeholders said, could be key as hospitals and banks rarely share data. They make little money from it and risk fines of up to ₹250 crore under India’s data protection law, if it leaks.

India’s existing consent-based data-sharing framework “solves the individual consent problem brilliantly, but it is not able to solve the collective problem,” a banking executive said, citing fraud that spans multiple banks as an example.

For now, Depa’s day-to-day management is expected to be handled by Samyog, a new industry group “powered by iSpirt.” Its website says it will register participants, provide standard contracts, and resolve disputes. Its board includes Sharma, and several leaders helped write the handbook. The site does not say Samyog is government-appointed.

Executives at iSpirt said that data owners, like hospitals and banks, would be paid each time their data is used. But, the handbook only says that payment methods “will be explored.” It also does not specify who would be liable if data were to leak from a sealed room.

The framework is not legally binding, and officials did not say if or when it might be. The path forward includes pilots in finance and healthcare and regulatory sandboxes where companies test models under supervision. Binding rules are more likely to come from sectoral regulators than through a new law, as was the case with India’s account aggregator system, which took five years to move from RBI rule-making to launch.

Much also depends on the Digital Personal Data Protection Act, 2023, whose main rules take effect in May 2027. India also continues to explore a dedicated AI legislation, while also debating if AI platforms need a completely new definition of how they are protected under law, Mint reported Wednesday.